[
https://issues.apache.org/jira/browse/KAFKA-9959?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Cheng Tan updated KAFKA-9959:
-----------------------------
Issue Type: Improvement (was: Bug)
> leastLoadedNode() does not provide a node fairly
> ------------------------------------------------
>
> Key: KAFKA-9959
> URL: https://issues.apache.org/jira/browse/KAFKA-9959
> Project: Kafka
> Issue Type: Improvement
> Reporter: Cheng Tan
> Assignee: Cheng Tan
> Priority: Major
>
> Currently, leastLoadedNode() provides a node with the following criteria:
> # Provide the connected node with least number of inflight requests
> # If no connected node exists, provide the connecting node with the largest
> index in the cached list of nodes.
> # If no connected or connecting node exists, provide the disconnected node
> which respects the reconnect backoff with the largest index in the cached
> list of nodes.
> However, criteria 2 and 3 may cause issues.
>
> Criteria 2: Since the timeoutCallsToSend() does not change the connection
> status, the node will remain a connecting status after the request time out.
> If no connected node exists, leastLoadedNode() will provide this same node
> until the socket timeout reached. It would be better to overlook the
> connecting node if any request has timed out on it.
>
> Criteria3: If the time interval between two invokes of leastLoadedNode() is
> greater than the reconnect.backoff.ms, the same disconnected node may be
> provided. We also want to pick a node with the least number of failed times.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
